JP2780458B2 - Vector quantization method and speech coding / decoding device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantization method and a speech encoding / decoding device used for a speech signal encoding / decoding device.

2. Description of the Related Art In the current field of digital mobile communication, in order to cope with an increase in the number of subscribers, a speech encoding / decoding apparatus capable of transmitting speech at a lower bit rate is desired. In order to perform the low bit rate audio coding, a large amount of information compression must be performed by utilizing the redundancy of the audio. As one of the information compression methods, a method called vector quantization has been conventionally used.

Vector quantization uses the statistical bias of a vector set when a plurality of pieces of analog value information are used as vectors,
The encoding is performed using a codebook characterized by storing a plurality of vectors (centroids). The vector to be coded is coded to the number of the closest centroid after calculating the distance to the codebook centroid. According to this method, a plurality of pieces of information are converted into one number, so that significant information compression is realized.

Problems to be Solved by the Invention However, the algorithm using the vector quantization requires a very large amount of calculation for calculating the distance from the centroid, so that there is a problem that real-time processing cannot be performed. In particular, in the compression of audio information having a large vector order, the amount of calculation becomes enormous. For example, if the degree of the vector is 100 and the codebook size is 2048, 2
It requires 0,000 multiplications and 2,000 comparisons. Assuming that the analysis unit of the digital audio signal is 20 msec, multiplication of 10 million times per second is required only by referring to the codebook.

According to the present invention, when the codebook is referred to, the calculation amount is exponentially reduced by applying the equation of the discrimination boundary surface between the centroid and the centroid stepwise without using the distance calculation with the centroid. It is going to reduce it greatly.

Means for Solving the Problems In order to achieve this object, a vector quantization method according to the present invention uses a set of centroids, which are vectors distributed on a multidimensional Euclidean space, and a set of centroids.
Providing a multidimensional plane coefficient storage unit that stores the coefficients of the multidimensional linear expression representing the plane to be divided, and substituting the elements of the input vector into the multidimensional linear expression composed of the coefficients stored in the multidimensional plane coefficient storage unit By narrowing the candidate set by referring to the sign of the value obtained by performing the search, the nearest centroid in the set of centroids is searched for the input vector.

Further, the speech encoding / decoding apparatus according to the present invention includes a framework encoding means for encoding several types of frameworks obtained by analyzing a speech signal with information of power and pitch, and a plurality of encoded inter-bone waveforms. An interosseous waveform codebook in which a centroid is stored, and a set of interosseous waveform centroids obtained by performing a two-division cluster analysis on the set of interosseous waveform centroids in the interosseous waveform codebook In order to vector-quantize the interskeletal waveform stretched between the frames obtained by the frame encoding means, and a multidimensional plane coefficient storage unit storing coefficients of a multidimensional linear expression representing a plane to be divided into two,
By narrowing the candidate set by referring to the sign of the value obtained by substituting the elements of the input vector into the multidimensional linear expression constituted by the coefficients stored in the multidimensional plane coefficient storage unit, An encoder having an interosseous waveform encoding means for selecting an interosseous waveform centroid closest to the interosseous waveform stretched between the set skeletons, and information based on the information encoded by the interosseous waveform encoding means. Skeleton decoding means for creating two types of skeletons, an interosseous waveform codebook in which a plurality of encoded interosseous waveform centroids are stored, and information encoded by the interosseous waveform encoding means. Originally, a decoder having an interstitial waveform decoding means for decoding an interstitial waveform stretched between the frames by using the interstitial waveform codebook is provided.

The present invention focuses on the fact that most of the calculation amount necessary for referring to the codebook is the calculation amount of the distance from the centroid. Substituting the centroid set into a multidimensional plane formula that divides the centroid set into two without calculating the distance to the Lloyd, finds out which set has the closest centroid from its positive or negative, and uses that information to determine the stage The candidate set is gradually reduced, and finally the closest centroid is obtained. As a result, the amount of calculation can be significantly reduced exponentially.

Embodiment To explain an embodiment of the present invention, as an example, first consider a problem of determining which of two centroids an input vector is closer to.

Let the order of the vector be d-th order. In the conventional method, since the distance between two centroids is determined to determine which is smaller, 2d multiplications by the Euclidean distance are required. Here, according to the method of substituting an input vector into an expression of a multidimensional plane equidistant from two centroids and determining the sign thereof, exactly the same performance can be obtained by d times of multiplication.

Therefore, assuming that the codebook has a size of 2K, a multidimensional plane formula for dividing the centroid set into two is prepared in stages, and the elements of the input vector are substituted into those formulas. By narrowing down the candidate set containing the closest centroid based on the sign of the obtained value and performing a search that finally selects one centroid, the calculation effect is reduced to 1 og ₂ K / K by the same effect Is done.

For example, if the degree of the vector is 100 and the codebook size is 2048, 200,000 multiplications are reduced to 1200.
Therefore, a great reduction in calculation can be realized, and the above object is achieved.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

As an embodiment of the present invention, an example of application to a speech encoding / decoding system using a skeleton (time series of pulses) and an interosseous waveform, which is one of speech encoding methods, will be described. FIG. 1 is a functional block connection diagram of a speech encoding / decoding system based on a skeleton and a waveform between bones according to an embodiment of the present invention.

In FIG. 1, 1 is an encoder, 2 is a decoder, 3 is an input voice, 4 is a power calculator, 5 is a pitch analyzer, 6 is a skeleton searcher, 7 is an interosseous waveform selector, and 8 is an interosseous. Waveform codebook, 9
Is a multidimensional plane coefficient storage unit, 10 is a skeleton forming unit, 11 is a waveform synthesizing unit, 12 is an interbone waveform codebook, and 13 is an output voice.

 The operation of each block will be described below.

In the encoder 1, first, a voice signal is sampled and converted into a digital signal, and a power calculation unit 4 converts the input voice divided for each fixed time length (one frame).
In, the average power in the section is determined and used as auxiliary information.

Next, in the pitch analysis unit 5, the pitch in the section is obtained and used as pitch information. In the skeleton search unit 6,
Based on the pitch information, a signal within one pitch is searched for two signals having the maximum absolute value in positive and negative directions, and the positions of the two signals, the signal amplitude, and the pitch information described above are skeleton information. And

The function of the interbone waveform selector 7 will be described with reference to FIG.
First, based on the skeleton information described above, within one pitch,
The waveform stretched between the positive maximum signal and the negative maximum signal, which is the skeleton, and the waveform stretched between the negative maximum signal and the positive maximum signal are obtained, and this is referred to as a basic interosseous waveform vector. I do.

Next, after two basic interosseous waveform vectors are normalized to fixed end points (temporally / power-wise), encoding is performed by vector quantization. Compared with the numbered interosseous waveform centroid stored in the interosseous waveform codebook 8, the number attached to the interosseous waveform centroid closest to the normalized basic interosseous waveform vector is the interosseous waveform. Information. In the present invention, the centroid selection is not performed by calculating the Euclidean distance with the interbone waveform centroid, but the centroid selection is performed by narrowing down the candidate set using the multidimensional plane coefficients stored in the multidimensional plane coefficient storage unit 9. .

Then, the power information, the skeleton information, and the interosseous waveform information are transmitted as a code of a unit time voice.

Here, an algorithm of the centroid selection by narrowing down the candidate set using the multidimensional plane coefficients stored in the multidimensional plane coefficient storage unit 9 in the interbone waveform selection unit 7 will be described below. FIG. 3 shows how the centroid is selected for a codebook size of 8.

<1> First, an input vector is substituted into an expression of a multidimensional plane that divides the entire set of centroids into two, and it is determined which of the sets is closer to the centroid from the positive or negative (FIG. 3 (a)). ).

<2> Next, the determined set is substituted into an equation of a multidimensional plane which divides the set into two, and it is determined which of the sets is closer to a centroid from the positive or negative (FIG. 3 (b)).

<3><2> is repeated until the candidate set has two elements (FIG. 3 (c)).

<4> Lastly, the centroid is substituted into an equation of a multidimensional plane equidistant from the two centroids, which one of the two centroids is closer to the other is determined, and the number of the centroid determined to be closer is calculated as a bone number. It is set as inter-waveform information (FIG. 3 (d)).

Also, the coefficient of the multidimensional plane equation used at this time is
Both are stored in the multidimensional plane coefficient storage unit 9,
The multidimensional plane coefficient storage unit 9 and the interbone waveform codebook 8 are created in the following procedure.

<1> A basic interosseous waveform vector obtained by analyzing speech in advance is collected for a large amount of speech data, and each is normalized to an end point fixed (temporally and power-wise) to create an interosseous waveform population. .

<2> A cluster analysis is performed on the interbone waveform population to obtain a specified number of centroids, which is referred to as an interbone waveform codebook 8.

<3> The centroid set of the interbone waveform codebook 8 is subjected to two-part cluster analysis using a multidimensional plane. Further, a two-part cluster analysis is performed on each of the two divided centroid sets. Then, the analysis is performed until the number of elements of the cluster becomes one, and the coefficients of the multidimensional plane equation that divides the two clusters obtained in the process are collected and stored in the multidimensional plane coefficient storage unit 9.

 Next, the operation of the decoder will be described with reference to FIG.

First, the skeleton forming unit 10 forms a voice skeleton based on the power information and the skeleton information obtained by the above-described encoding. The upper part of FIG. 4 is an example of this skeleton,
It shows a state where a skeleton is formed based on skeleton information.
The waveform synthesizing section 11 selects a basic interosseous waveform from the same interosseous waveform codebook 12 stored in the encoder 1 on the basis of the interosseous waveform information, and temporally and power-wise according to the skeleton. The waveform is converted and stretched between the bones, and this waveform is used as an output voice. The lower part of FIG. 4 shows an example of this waveform synthesis. This shows a state in which a basic interosseous waveform is stretched between frames by using two types of interosseous waveform samples selected from the interosseous waveform codebook 12 based on the interosseous waveform information.

In order to demonstrate the effects of the present invention, a discriminative simulation experiment using a random codebook was performed. As the centroid set, a set of random number vectors of degree 10 is used. As a sample set to be input, a set of random number vectors different from the centroid set is used. In each vector, each value of the vector is a random number from 0 to 1, and the size of the set is 1024. The conventional method using the Euclidean distance is referred to as method 1, and the search method using the multidimensional plane according to the present invention is referred to as method 2.
Perform a search as In addition, a method of selecting 32 centroids from the vicinity of the centroid searched by the method 2 and obtaining the Euclidean distance from the candidates to obtain the nearest centroid is referred to as a method 3. Method 3 is an improvement over method 2. When searching on a multidimensional plane, the centroid with the shortest distance is not always selected. Therefore, in the case of the methods 2 and 3, the performance of the method depends on how small the centroid can be searched. In order to evaluate the performance, the average number of the selected centroids in the order of the Euclidean distance and the number of the selected centroids are calculated.

Table 1 on page 15 of this specification shows that 1024
It shows the number of multiplications for the number of samples, the average rank, and the number selected as the first place.

This experiment has verified that the present invention can achieve a significant reduction in the amount of calculation with almost the same performance.

Effects of the Invention As described above, the present invention substitutes the centroid set into a multidimensional plane equation for dividing the centroid set into two without calculating the distance to the centroid to find the nearest centroid, and calculates The set of nearby centroids is determined, the centroid candidate set is reduced in stages, and finally the closest centroid is determined, so the computational cost is significantly reduced exponentially can do.

[Brief description of the drawings]

FIG. 1 is a block diagram of a vector quantization method and a speech coding / decoding apparatus according to an embodiment of the present invention. FIG. 2 is a conceptual diagram showing the operation of the interosseous waveform selector 7 in the embodiment. 3 (a) to 3 (d) are conceptual diagrams showing a centroid selecting operation of the interosseous waveform selecting unit 7 in the embodiment, and FIG. 4 is a conceptual diagram showing an operation of the decoder 2 in the embodiment. It is. 1 ... encoder, 2 ... decoder, 3 ... input voice, 4 ...
Power calculation section, 5: Pitch analysis section, 6: Frame search section, 7: Interbone waveform selection section, 8: Interbone waveform codebook, 9
... Multidimensional plane coefficient storage unit, 10 ... Frame forming unit, 11 ...
Waveform synthesizing unit, 12: interstitial waveform codebook, 13: output voice.

Claims (2)

(57) [Claims] 1. A multidimensional plane coefficient storage unit storing a set of centroids, which are vectors distributed on a multidimensional Euclidean space, and a multidimensional linear coefficient representing a plane dividing the centroid set into two. By narrowing down the candidate set with reference to the sign of the value obtained by substituting the elements of the input vector into the multidimensional linear expression constituted by the coefficients stored in the multidimensional plane coefficient storage unit, A vector quantization method for searching a centroid closest to the set of vectors from the set of centroids. 2. A skeleton encoding means for encoding several types of skeletons obtained by analyzing an audio signal with power and pitch information, and a skeleton storing a plurality of encoded interskeletal waveform centroids. Representing a plane which divides the set of interstitial waveform centroids obtained by performing a two-division cluster analysis on the set of interstitial waveform centroids in the interstitial waveform codebook A multidimensional plane coefficient storage unit in which coefficients of a linear expression are stored, and the multidimensional plane coefficient storage unit for vector quantizing an interosseous waveform stretched between frames obtained by the interosseous encoding means. By substituting the elements of the input vector into the linear equation composed of the coefficients stored in the section, the candidate set is narrowed down by referring to the sign of the value obtained by the skeleton encoding means. An encoder having an interosseous waveform encoding means for selecting an interosseous waveform centroid closest to the interosseous waveform stretched between the skeletons, and information based on the information encoded by the interosseous waveform encoding means. Skeleton decoding means for creating two types of skeletons, an interosseous waveform codebook in which a plurality of encoded interosseous waveform centroids are stored, and coded by the interosseous waveform encoding means. Based on the information
A speech encoding / decoding apparatus, comprising: a decoder having an interosseous waveform decoding means for decoding an interosseous waveform stretched between frames by using the interosseous waveform codebook.